mssfj/qwen25-openmathinstruct2-sft-50k

Qwen2.5 OpenMathInstruct-2 SFT 50K

A 50,000-example, short-context mathematics SFT dataset extracted from NVIDIA OpenMathInstruct-2. It was prepared for instruction tuning mssfj/qwen25-0.5b-finemath-4plus and subsequent GSM8K and MATH-500 benchmarking, including controlled comparisons with TDT. This is an extracted and filtered dataset, not a newly generated set of solutions.

Composition

MATH-related examples make up 70% of each split; GSM8K-related examples make up 30%. These categories describe the upstream problem_source, not a newly assigned difficulty label. Each selected numeric problem template occurs only once across the entire dataset.

Split	Examples	MATH-related	GSM8K-related	Total formatted tokens	Maximum tokens
train	49,000	34,300	14,700	21,212,482	1854
validation	1,000	700	300	427,265	1868

The token limit is 2,048 per complete conversation, including the system prompt, user problem, full assistant solution, and Qwen chat-template delimiters. Problems and solutions are not truncated. The published solution text is retained apart from trimming outer whitespace; no answer line or solution was regenerated or appended.

Source and provenance

• Source: nvidia/OpenMathInstruct-2, split train_1M.
• Source revision: 469216e3f46f4dacf476b382e192485ea51a143e.
• Upstream fields: problem, generated_solution, expected_answer, problem_source.
• Upstream teacher model: Llama3.1-405B-Instruct, as documented by NVIDIA.
• Tokenizer: mssfj/qwen25-0.5b-finemath-4plus, revision 0600e6f3b418e77d7323b8934451ac66d49a1185, using its Qwen2.5 tokenizer and chat template.
• Selection and split seed: 42.

The source paper is OpenMathInstruct-2: Accelerating AI for Math with Massive Open-Source Instruction Data, by Shubham Toshniwal, Wei Du, Ivan Moshkov, Branislav Kisacanin, Alexan Ayrapetyan, and Igor Gitman (2024).

Selection procedure

1. Scan all 1,000,000 rows in the pinned train_1M files in sorted filename order.
2. Reject empty required fields, unknown source categories, problems containing [asy] diagram markup, and problem/solution text containing chat control-token strings.
3. Require a balanced final \boxed{...} or \fbox{...} and compare it with expected_answer after limited whitespace and LaTeX-format normalization. This is a string-consistency check, not verification of the mathematical reasoning.
4. Deduplicate normalized problems. Choose the shortest answer-consistent solution by character length, with a seeded SHA-256 tie-breaker.
5. Render complete conversations with the pinned chat template and reject lengths greater than 2,048 tokenizer tokens.
6. Deduplicate numeric templates. Normalization uses NFKC, lowercase text, removes selected LaTeX presentation commands, and keeps letter, number, and math-operator tokens. Numeric templates replace numeric literals with NUM. Keep the problem with the lowest seeded SHA-256 rank per template.
7. Form a deterministic candidate pool of up to 70,000 MATH-related and 30,000 GSM8K-related examples. Apply the benchmark checks below to this candidate pool.
8. Select 35,000 MATH-related and 15,000 GSM8K-related surviving examples by seeded hash. Assign 700 MATH-related and 300 GSM8K-related examples to validation using a separate seeded hash ordering, leaving 49,000 training examples. Shuffle file order by another deterministic hash.
9. Retokenize every exported example to verify its saved length and check that normalized problem IDs and numeric-template group IDs are disjoint between splits.

This favors short complete solutions and template diversity. It is not a representative random sample of the original corpus or its difficulty distribution. Derived problems whose templates change substantially may still occur in both train and validation.

Benchmark overlap filtering

Reference sets:

• openai/gsm8k, main, test: 1,319 problems, revision 740312add88f781978c0658806c59bc2815b9866.
• HuggingFaceH4/MATH-500, test: 500 problems, revision 6e4ed1a2a79af7d8630a6b768ec859cb5af4d3be.

An example is removed when its problem meets any of these conditions:

• Exact match after problem normalization.
• Exact numeric-template match, targeting number-substitution variants.
• RapidFuzz normalized-problem ratio of at least 90 against any benchmark problem.
• Maximum sentence-embedding cosine similarity of at least 0.94.
• Maximum cosine similarity of at least 0.90, together with a normalized-problem ratio of at least 70 or a numeric-template ratio of at least 85 for that nearest benchmark.

Embedding model: sentence-transformers/all-MiniLM-L6-v2, revision 1110a243fdf4706b3f48f1d95db1a4f5529b4d41, with normalized embeddings and a maximum input length of 512 embedding-model tokens. Longer questions are truncated for this similarity check only, not in the SFT data. These thresholds are heuristics: they can remove related but distinct problems and miss paraphrases or mathematical equivalents. This dataset does not guarantee complete semantic decontamination.

The filtering checks problem text, not all mentions within generated solutions. It also does not undo possible benchmark exposure during the base model's prior FineWeb-Edu or FineMath pretraining, or during the teacher model's training. GSM8K training problems remain eligible; only the named evaluation sets are exclusion references. MATH test problems outside MATH-500 were not an additional exclusion set.

decontamination_audit.jsonl records removed candidate IDs, matched benchmark row IDs, and matching scores. It does not reproduce benchmark questions or answers. selection_stats.json records filter counts, split composition, source revisions, template/script fingerprints, token statistics, and Parquet SHA-256 checksums.

Fields

Field	Description
messages	System/user/assistant messages for SFT
problem	Selected upstream problem
generated_solution	Selected upstream solution
expected_answer	Upstream reference or majority-vote answer
problem_source	math, augmented_math, gsm8k, or augmented_gsm8k
problem_id	SHA-256 of the normalized problem
problem_group_id	SHA-256 of its numeric template
token_count	Complete conversation length under the pinned tokenizer/template
source_dataset, source_revision, source_split	Upstream provenance
source_file, source_row	Source Parquet path and zero-based row number
nearest_benchmark_cosine	Maximum embedding similarity to the exclusion references; audit metadata only

Usage

from datasets import load_dataset

dataset = load_dataset("mssfj/qwen25-openmathinstruct2-sft-50k")
train = dataset["train"]
validation = dataset["validation"]
print(train[0]["messages"])

Use messages for training and keep the reference/audit fields out of the prompt. The system prompt is:

Solve the math problem step by step. Put the final answer in \boxed{...}.

Render with the pinned Qwen chat template (chat_template.jinja is also supplied), train with loss on the assistant response and its turn-ending token, and keep the input embeddings and output weights untied to match the base checkpoint. At inference, stop on the Qwen assistant turn terminator <|im_end|> as well as the base EOS token. Use this dataset's validation split for model selection; keep GSM8K test and MATH-500 for final evaluation. For TDT comparisons, fix the dataset revision and examples, initial weights, training token budget, template, and evaluation settings.

Reproduction

The included prepare_openmath_sft.py downloads pinned input revisions and rebuilds the Parquet files and selection statistics. A CUDA GPU accelerates embedding checks. Install a suitable PyTorch build for your system, then:

pip install -r requirements.txt
python prepare_openmath_sft.py \
  --output ./rebuilt-openmath-sft-50k \
  --total 50000 --validation 1000 --math-ratio 0.70 \
  --max-tokens 2048 --seed 42 --candidate-multiplier 2

environment.json records the generation environment, including the CUDA PyTorch build. Different numerical environments can change borderline embedding decisions. Compare output checksums and statistics when reproducing the dataset. The build script creates data and statistics; this README and dependency/environment files accompany the published release separately.

Build statistics

Counters refer to the stage where each filter is applied; they are not all counts over the same population. Similarity checks apply to the hash-selected candidate pool.

Counter	Count
source_rows	1,000,000
diagram_or_control_token	21,843
missing_or_mismatched_final_box	993
duplicate_normalized_problem	386,845
empty_field	733
unique_answer_checked_candidates	589,586
over_token_limit	355
duplicate_numeric_template	28,173
eligible_math	481,746
eligible_gsm8k	79,312
similarity_candidate_pool	100,000
benchmark_fuzzy_ratio_90	81
benchmark_numeric_template_exact	11
embedding_input_over_512_tokens	169
benchmark_semantic_similarity	11

License and attribution

This derived dataset follows the upstream OpenMathInstruct-2 dataset's CC BY 4.0 license. Attribute NVIDIA and the OpenMathInstruct-2 authors, and identify this release as an extracted subset with formatting, consistency, length, deduplication, benchmark overlap filtering, and split changes. Refer to the upstream dataset card for the original release and citation.